[lkml]   [2015]   [Apr]   [21]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [RFC][PATCHSET] non-recursive link_path_walk() and reducing stack footprint
On Mon, Apr 20, 2015 at 07:12:22PM +0100, Al Viro wrote:

> Patches 2/24..6/24 are from Neil's RCU follow_link patchset; the
> rest of his patchset is, of course, derailed by the massage done here,
> but AFAICS we should be able to port it on top of this one with reasonably
> little PITA.

BTW, looking at the ->put_link() instances in the tree, after this series
all but one of them ignore *everything* other than cookie. The only exception
is hppfs; it wants dentry (and its inode as well):

static void hppfs_put_link(struct dentry *dentry, void *cookie)
struct dentry *proc_dentry = HPPFS_I(d_inode(dentry))->proc_dentry;

if (d_inode(proc_dentry)->i_op->put_link)
d_inode(proc_dentry)->i_op->put_link(proc_dentry, cookie);

Does anybody have objections against passing them inodes instead of dentries?
It would be a lot more convenient for RCU purposes...

Other observations about the Neil's series:
* the games played in page_follow_link_light/page_put_link are
completely pointless; this
void page_put_link(struct dentry *dentry, char *link, void *cookie)
- struct page *page = cookie;
+ struct page *page = (void *)((unsigned long)cookie & ~1UL);

if (page) {
- kunmap(page);
+ if (page == cookie)
+ kunmap(page);
is trivially avoided by making sure that inodes with such ->put_link() in
their inode_operations have
mapping_get_gfp_mask(mapping) & ~__GFP_HIGHMEM);
done to them. And to hell with that kmap/kunmap pair - it's a bad idea to
have kmap held across practically unlimited period, anyway (run into a hung
NFS server while traversing the symlink body, get stuck for minutes until it
times out). I'll take care of that, it's not much work.
* XFS: is there any reason why you guys don't separate inline and
non-inline symlinks? Or don't just use page_follow_link_light() for the
latter... Note, BTW, that NUL-termination for the latter will be taken
care of by page_getlink(). It's inline ones that look like crap, actually -
unless I'm misreading that code, the longest possible inline symlink will
have ip->i_df.if_u1.if_data completely filled, with no place for terminating
NUL. Is that correct? If so, it might make sense to have three variants -
short (== inline shorter than XFS_IFORK_DSIZE) just having ->follow_link()
return ip->i_df.if_u1.if_data and have xfs_setup_inode() do nd_terminate_link()
to make sure they are NUL-terminated, long (non-inline) just using
page_follow_link_light() and bogus (inline with length exactly equal to
XFS_IFORK_DSIZE); the last would be the only ones with XFS-specific non-trivial
->follow_link() - those really need to allocate a buffer and copy into it...

FWIW, there's a potentially interesting alternative - the thing is,
normal callers of link_path_walk() already know the length of name; both
getname() and getname_kernel() calculate the lengths. And ->follow_link()
tend to have the lengths available as well (and usually equal to ->i_size of
the inode in question). So we might consider passing (pointer,length) pairs
to link_path_walk(). In principle, it might simplify things... Comments?

* logfs has ->follow_link equal to page_follow_link_light and
NULL ->put_link. Obvious leak, and that one's -stable fodder - it had
been there all way back to original merge. I'll send a fix in a minute.

 \ /
  Last update: 2015-04-21 17:01    [W:0.150 / U:0.284 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site